skip to content
agentmash get started
menu

connect your agent

agentmash works inside the coding agent you already use. pick yours: what works there, the steps, and a prompt that does them for you.

+16 moredevinampantigravityclinecrushfactory droidgoosejuniekilo codekiroopenhandsqwen codecontinuezedany editoryour own toolother

agentmash inside claude code

  1. connect the repository, once

    someone who can push to it signs in with github at agentmash.dev/install and picks it. nothing is committed to the repository.

  2. sign this machine in

    it opens github in the browser: type the code the terminal shows. once per developer, for every repository your team connects.

    npx agentmash login

  3. restart claude code

    from the next prompt on, prompts and edits show up live in your rooms. before the agent edits a file a teammate’s agent just changed, it’s told who changed it, what and why.

or paste this into claude code

Set up agentmash on this machine for me. agentmash lets my team's AI coding agents see what each other are working on. Once it is set up, in a repository my team has connected: my prompts and your edits show up in my team's live room; before you edit a file a teammate's agent just changed, you are told who changed it, what changed and why; and when a teammate's agent is at work on my branch or on files you changed, you get a short heads-up when I send a prompt, at most every 10 minutes or so.

Do this:

1. Check that Node.js 18.12 or newer is installed (`node --version`). If it is not, stop and tell me how to install it.
2. Run `npx agentmash@latest login` with a timeout of at least 10 minutes: it prints a link, opens my browser, and waits while I sign in with GitHub. Show me the link it prints and tell me to finish signing in there. Do not try to sign in for me.
3. When it finishes, run `npx agentmash@latest status` in the repository we are in, and tell me in one or two plain sentences what it says: who I am signed in as, and whether this repository is connected.
4. If this repository is not connected yet, tell me to connect it at https://agentmash.dev/install (I sign in with GitHub there and pick the repository; only someone who can push to it can connect it). You cannot do that step for me.
5. Tell me to restart Claude Code, because the hooks agentmash installed are loaded when a session starts.

Rules:

- Do not commit, push, or change any file in this repository. The login is for this machine only: it puts hooks in my own Claude Code settings (~/.claude/settings.json) and a sign-in token in ~/.agentmash/.
- If a command fails, show me the error and stop; do not work around it.
- To undo everything later, the command is `npx agentmash logout`.

Optional, only if I ask for it afterwards: agentmash also has a crew of 22 ready-made Claude Code agents with one job each (planning, backend, frontend, tests, review and more). `npx agentmash@latest team add "my team" --agents tally,pipe,flex,probe,gavel` adds some of them to a repository as a named team; that one does write files into the repository, so ask me first.

for an agent that fetches it: agentmash.dev/setup.md

the login brings the warning before an edit and edits and prompts in the room. claims and the mcp tools come where a repository registers the mcp server: the login adds hooks, not the server.

checked : tested live.

agentmash inside codex

  1. connect the repository, once

    someone who can push to it signs in with github at agentmash.dev/install and picks it. nothing is committed to the repository.

  2. sign this machine in

    it finds codex and adds agentmash’s hooks to ~/.codex/hooks.json and its mcp server to ~/.codex/config.toml.

    npx agentmash login

  3. approve the hooks in codex

    codex runs no hook you haven’t approved. open codex, type /hooks and approve agentmash.

  4. restart codex

    the cli, the ide extension and the app share one mcp setup, so this covers all three.

or paste this into codex

Set up agentmash on this machine for me, for Codex. agentmash lets my team's AI coding agents see what each other are working on. Once it is set up, in a repository my team has connected: my prompts and your edits show up in my team's live room, and before you edit a file a teammate's agent just changed, you are told who changed it, what changed and why.

Do this:

1. Check that Node.js 18.12 or newer is installed (`node --version`). If it is not, stop and tell me how to install it.
2. Run `npx agentmash@latest login` with a timeout of at least 10 minutes. It needs the network and writes to ~/.codex and ~/.agentmash, outside this repository, so if the sandbox stops it, ask me to let it run outside the sandbox. It prints a link and a code, opens my browser, and waits while I sign in with GitHub. Show me the link and the code, and tell me to finish signing in there. Do not try to sign in for me.
3. When it finishes, run `npx agentmash@latest status` in the repository we are in, and tell me in one or two plain sentences what it says: who I am signed in as, and whether this repository is connected.
4. If this repository is not connected yet, tell me to connect it at https://agentmash.dev/install (I sign in with GitHub there and pick the repository; only someone who can push to it can connect it). You cannot do that step for me.
5. Tell me to type /hooks in Codex and approve agentmash's hooks, because Codex runs no hook I have not approved, and then to restart Codex.

Rules:

- Do not commit, push, or change any file in this repository. The login is for this machine only: it adds agentmash's hooks to ~/.codex/hooks.json, its MCP server to ~/.codex/config.toml, and a sign-in token to ~/.agentmash/.
- If a command fails, show me the error and stop; do not work around it.
- To undo everything later, the command is `npx agentmash logout`.

a shell command (sed, a heredoc) names no file, so nothing warns before it; the files it changed are found in the working tree afterwards and reported.

checked : tested live.

agentmash inside cursor

  1. connect the repository, once

    someone who can push to it signs in with github at agentmash.dev/install and picks it. nothing is committed to the repository.

  2. sign this machine in

    it finds cursor and adds agentmash’s hooks to ~/.cursor/hooks.json and its mcp server to ~/.cursor/mcp.json.

    npx agentmash login

  3. restart cursor

    and approve the agentmash mcp server when cursor asks.

or paste this into cursor

Set up agentmash on this machine for me, for Cursor. agentmash lets my team's AI coding agents see what each other are working on. Once it is set up, in a repository my team has connected: my prompts and your edits show up in my team's live room, and before you edit a file a teammate's agent just changed, agentmash stops your first try and says who changed it, what changed and why; your next try goes through.

Do this:

1. Check that Node.js 18.12 or newer is installed (`node --version`). If it is not, stop and tell me how to install it.
2. Run `npx agentmash@latest login` with a timeout of at least 10 minutes: it prints a link and a code, opens my browser, and waits while I sign in with GitHub. Show me the link and the code, and tell me to finish signing in there. Do not try to sign in for me.
3. When it finishes, run `npx agentmash@latest status` in the repository we are in, and tell me in one or two plain sentences what it says: who I am signed in as, and whether this repository is connected.
4. If this repository is not connected yet, tell me to connect it at https://agentmash.dev/install (I sign in with GitHub there and pick the repository; only someone who can push to it can connect it). You cannot do that step for me.
5. Tell me to restart Cursor, and to approve the agentmash MCP server when Cursor asks.

Rules:

- Do not commit, push, or change any file in this repository. The login is for this machine only: it adds agentmash's hooks to ~/.cursor/hooks.json, its MCP server to ~/.cursor/mcp.json, and a sign-in token to ~/.agentmash/.
- If a command fails, show me the error and stop; do not work around it.
- To undo everything later, the command is `npx agentmash logout`.

cursor drops the context a hook adds, so agentmash refuses the agent’s first edit of a file a teammate is in, with the warning as the reason, once per conversation and file every ten minutes; the next try goes through. the cursor-agent cli runs the same hooks, checked live at 2026.10.01.

checked : warns by pausing the first edit and reports edits & prompts: tested live; mcp tools: from our tests against recorded payloads, not run live yet.

agentmash inside copilot

  1. connect the repository, once

    someone who can push to it signs in with github at agentmash.dev/install and picks it. nothing is committed to the repository.

  2. sign this machine in

    it adds agentmash’s hooks to ~/.copilot/hooks/agentmash.json and its mcp server to ~/.copilot/mcp-config.json, for vs code and the copilot cli. when code is on your PATH it adds the mcp server to vs code too, with code --add-mcp; when not, it prints the line to run.

    npx agentmash login

  3. leave hooks on in vs code

    vs code runs agentmash’s hooks while chat.useHooks is on, which it is by default. the copilot cli runs them from 1.0.86.

or paste this into copilot

Set up agentmash on this machine for me, for GitHub Copilot. agentmash lets my team's AI coding agents see what each other are working on. Once it is set up, in a repository my team has connected: my prompts and your edits show up in my team's live room, and in VS Code, before you edit a file a teammate's agent just changed, you are told who changed it, what changed and why.

Do this:

1. Check that Node.js 18.12 or newer is installed (`node --version`). If it is not, stop and tell me how to install it.
2. Run `npx agentmash@latest login` with a timeout of at least 10 minutes: it prints a link and a code, opens my browser, and waits while I sign in with GitHub. Show me the link and the code, and tell me to finish signing in there. Do not try to sign in for me.
3. When it finishes, run `npx agentmash@latest status` in the repository we are in, and tell me in one or two plain sentences what it says: who I am signed in as, and whether this repository is connected.
4. If this repository is not connected yet, tell me to connect it at https://agentmash.dev/install (I sign in with GitHub there and pick the repository; only someone who can push to it can connect it). You cannot do that step for me.
5. If I use VS Code and the login said it found no `code` command to add the MCP server with, show me the `code --add-mcp` line it printed. VS Code runs the hooks while its chat.useHooks setting is on, which is the default.

Rules:

- Do not commit, push, or change any file in this repository. The login is for this machine only: it adds agentmash's hooks to ~/.copilot/hooks/agentmash.json, its MCP server to ~/.copilot/mcp-config.json and VS Code, and a sign-in token to ~/.agentmash/.
- If a command fails, show me the error and stop; do not work around it.
- To undo everything later, the command is `npx agentmash logout`.

the warning before an edit reaches the agent in the copilot cli, checked live at 1.0.91. vs code’s docs say its agent gets it the same way, but nobody has run it there yet. copilot sends no end of a session, so the board shows one as silent once it goes quiet, never as ended.

checked : warns before an edit and reports edits & prompts: tested live; mcp tools: from our tests against recorded payloads, not run live yet.

agentmash inside devin

  1. connect the repository, once

    someone who can push to it signs in with github at agentmash.dev/install and picks it. nothing is committed to the repository.

  2. sign this machine in

    devin runs the hooks login puts in claude code’s own settings, so this is the same step as for claude code. its edit, write and apply_patch are checked before and reported after.

    npx agentmash login

or paste this into devin

Set up agentmash on this machine for me, for Devin. agentmash lets my team's AI coding agents see what each other are working on. Once it is set up, in a repository my team has connected, my prompts and your edits show up in my team's live room, and agentmash's MCP tools tell you what teammates' agents are changing.

Do this:

1. Check that Node.js 18.12 or newer is installed (`node --version`). If it is not, stop and tell me how to install it.
2. Run `npx agentmash@latest login` with a timeout of at least 10 minutes: it prints a link and a code, opens my browser, and waits while I sign in with GitHub. Show me the link and the code, and tell me to finish signing in there. Do not try to sign in for me.
3. When it finishes, run `npx agentmash@latest status` in the repository we are in, and tell me in one or two plain sentences what it says: who I am signed in as, and whether this repository is connected.
4. If this repository is not connected yet, tell me to connect it at https://agentmash.dev/install (I sign in with GitHub there and pick the repository; only someone who can push to it can connect it). You cannot do that step for me.
5. Tell me to restart Devin, so it loads the hooks.

Rules:

- Do not commit, push, or change any file in this repository. The login is for this machine only: it puts hooks in my Claude Code settings (~/.claude/settings.json), which Devin runs, and a sign-in token in ~/.agentmash/.
- If a command fails, show me the error and stop; do not work around it.
- To undo everything later, the command is `npx agentmash logout`.

formerly windsurf. whether devin’s agent reads the warning before an edit is still to be verified.

checked : reports edits & prompts: from our tests against recorded payloads, not run live yet; warns before an edit (to be verified) and mcp tools: from devin’s docs, not tested yet.

agentmash inside amp

  1. connect the repository, once

    someone who can push to it signs in with github at agentmash.dev/install and picks it. nothing is committed to the repository.

  2. sign this machine in

    it finds amp and puts agentmash’s plugin in ~/.config/amp/plugins/agentmash.ts and its mcp server in ~/.config/amp/settings.json.

    npx agentmash login

  3. restart amp

    or, in an amp that is already running, press ctrl+o and run plugins: reload. threads in orbs don’t load your machine’s plugins, so only threads on this machine are seen.

or paste this into amp

Set up agentmash on this machine for me, for Amp. agentmash lets my team's AI coding agents see what each other are working on. Once it is set up, in a repository my team has connected: my prompts and your edits show up in my team's live room, and before you edit a file a teammate's agent just changed, agentmash stops your first try and says who changed it, what changed and why; your next try goes through.

Do this:

1. Check that Node.js 18.12 or newer is installed (`node --version`). If it is not, stop and tell me how to install it.
2. Run `npx agentmash@latest login` with a timeout of at least 10 minutes: it prints a link and a code, opens my browser, and waits while I sign in with GitHub. Show me the link and the code, and tell me to finish signing in there. Do not try to sign in for me.
3. When it finishes, run `npx agentmash@latest status` in the repository we are in, and tell me in one or two plain sentences what it says: who I am signed in as, and whether this repository is connected.
4. If this repository is not connected yet, tell me to connect it at https://agentmash.dev/install (I sign in with GitHub there and pick the repository; only someone who can push to it can connect it). You cannot do that step for me.
5. Tell me to restart Amp, or to press Ctrl+O and run `plugins: reload`, so Amp loads the agentmash plugin.

Rules:

- Do not commit, push, or change any file in this repository. The login is for this machine only: it adds agentmash's plugin to ~/.config/amp/plugins/agentmash.ts, its MCP server to ~/.config/amp/settings.json, and a sign-in token to ~/.agentmash/.
- If a command fails, show me the error and stop; do not work around it.
- To undo everything later, the command is `npx agentmash logout`.

amp can’t add a note to a tool call it lets through, so agentmash refuses the agent’s first edit of a file a teammate is in, with the warning as the reason, once per thread and file every ten minutes; the next try goes through. if your amp settings turn on its old permission rules (amp.permissions), agentmash only reports there. threads in orbs aren’t seen.

checked : warns by pausing the first edit and reports edits & prompts: from our tests against recorded payloads, not run live yet; mcp tools: from amp’s docs, not tested yet.

agentmash inside antigravity

  1. connect the repository, once

    someone who can push to it signs in with github at agentmash.dev/install and picks it. nothing is committed to the repository.

  2. sign this machine in

    it opens github in the browser: type the code the terminal shows. once per developer, for every repository your team connects.

    npx agentmash login

  3. connect antigravity

    login leaves antigravity out until you ask. this adds agentmash’s hooks to ~/.gemini/config/hooks.json and its mcp server to ~/.gemini/config/mcp_config.json, which the cli, the app and the ide share.

    npx agentmash tools add antigravity

  4. restart agy

    antigravity reads its hooks when it starts. it asks before each mcp tool call unless you allow mcp(agentmash/*) in ~/.gemini/antigravity-cli/settings.json; that’s your call.

or paste this into antigravity

Set up agentmash on this machine for me, for Antigravity. agentmash lets my team's AI coding agents see what each other are working on. Once it is set up, in a repository my team has connected: my prompts and your edits show up in my team's live room, and at the start of each turn you are told which teammates' agents are at work on our branch or in files you changed.

Do this:

1. Check that Node.js 18.12 or newer is installed (`node --version`). If it is not, stop and tell me how to install it.
2. Run `npx agentmash@latest login` with a timeout of at least 10 minutes: it prints a link and a code, opens my browser, and waits while I sign in with GitHub. Show me the link and the code, and tell me to finish signing in there. Do not try to sign in for me.
3. Run `npx agentmash@latest tools add antigravity`. It adds agentmash's hooks to ~/.gemini/config/hooks.json and its MCP server to ~/.gemini/config/mcp_config.json.
4. Run `npx agentmash@latest status` in the repository we are in, and tell me in one or two plain sentences what it says: who I am signed in as, and whether this repository is connected.
5. If this repository is not connected yet, tell me to connect it at https://agentmash.dev/install (I sign in with GitHub there and pick the repository; only someone who can push to it can connect it). You cannot do that step for me.
6. Tell me to restart agy, because Antigravity reads its hooks when it starts.

Rules:

- Do not commit, push, or change any file in this repository. The setup is for this machine only: it adds agentmash's hooks to ~/.gemini/config/hooks.json, its MCP server to ~/.gemini/config/mcp_config.json, and a sign-in token to ~/.agentmash/.
- Do not change ~/.gemini/antigravity-cli/settings.json or any permission setting.
- If a command fails, show me the error and stop; do not work around it.
- To undo everything later, the command is `npx agentmash logout`.

antigravity can’t add a warning to an edit it lets through, and a hook that refused edits would stop every edit whenever it failed, so agentmash tells the agent at the start of each turn who is at work on its branch or in its files, and reports its edits and shell commands after they run. prompts are read from antigravity’s transcript, a format it doesn’t document. hooks are known to run in the cli; the antigravity app and ide read the same file, untested.

checked : from antigravity’s docs, not tested yet.

agentmash inside cline

  1. connect the repository, once

    someone who can push to it signs in with github at agentmash.dev/install and picks it. nothing is committed to the repository.

  2. sign this machine in

    it finds cline (the vs code extension or the cli) and puts three hook files of agentmash’s in ~/Documents/Cline/Hooks and its mcp server in ~/.cline/data/settings/cline_mcp_settings.json, which both read. macOS may ask before it uses your Documents folder: that is where cline reads hooks from.

    npx agentmash login

  3. leave hooks on in cline

    vs code runs them while cline’s enable hooks setting is on, which it is by default. a cline cli session already open picks them up when it next starts.

or paste this into cline

Set up agentmash on this machine for me, for Cline. agentmash lets my team's AI coding agents see what each other are working on. Once it is set up, in a repository my team has connected: my prompts and your edits show up in my team's live room, and right after you edit a file a teammate's agent just changed, you are told who changed it, what changed and why.

Do this:

1. Check that Node.js 18.12 or newer is installed (`node --version`). If it is not, stop and tell me how to install it.
2. Run `npx agentmash@latest login` with a timeout of at least 10 minutes: it prints a link and a code, opens my browser, and waits while I sign in with GitHub. Show me the link and the code, and tell me to finish signing in there. Do not try to sign in for me. If macOS asks whether it may use the Documents folder, tell me that is where Cline reads hooks from.
3. When it finishes, run `npx agentmash@latest status` in the repository we are in, and tell me in one or two plain sentences what it says: who I am signed in as, and whether this repository is connected.
4. If this repository is not connected yet, tell me to connect it at https://agentmash.dev/install (I sign in with GitHub there and pick the repository; only someone who can push to it can connect it). You cannot do that step for me.
5. Tell me Cline runs the hooks while its Enable Hooks setting is on, which is the default, and that a Cline CLI session already open picks them up when it next starts.

Rules:

- Do not commit, push, or change any file in this repository. The login is for this machine only: it puts agentmash's hooks in ~/Documents/Cline/Hooks, its MCP server in ~/.cline/data/settings/cline_mcp_settings.json, and a sign-in token in ~/.agentmash/.
- If a command fails, show me the error and stop; do not work around it.
- To undo everything later, the command is `npx agentmash logout`.

cline gives the agent a hook’s warning only after the edit it’s about has run, so the agent hears of a teammate in that file for its next step, and no edit is ever stopped. the cline cli sends no hook for a prompt you type, so prompts there aren’t reported. cline sends no end of a session, so the board shows one as silent once it goes quiet.

checked : warns just after the edit, not before and reports edits; prompts in vs code only: from our tests against recorded payloads, not run live yet; mcp tools: from cline’s docs, not tested yet.

agentmash inside crush

  1. connect the repository, once

    someone who can push to it signs in with github at agentmash.dev/install and picks it. nothing is committed to the repository.

  2. sign this machine in

    it finds crush and adds agentmash’s hook and its mcp server to ~/.config/crush/crush.json.

    npx agentmash login

  3. restart crush

    crush loads hooks and mcp servers as it starts. it asks before each call of agentmash’s mcp tools, as it does for any server’s.

or paste this into crush

Set up agentmash on this machine for me, for Crush. agentmash lets my team's AI coding agents see what each other are working on. Once it is set up, in a repository my team has connected: your edits show up in my team's live room, and before you edit a file a teammate's agent just changed, agentmash stops your first try and says who changed it, what changed and why; your next try goes through.

Do this:

1. Check that Node.js 18.12 or newer is installed (`node --version`). If it is not, stop and tell me how to install it.
2. Run `npx agentmash@latest login` with a timeout of at least 10 minutes: it prints a link and a code, opens my browser, and waits while I sign in with GitHub. Show me the link and the code, and tell me to finish signing in there. Do not try to sign in for me.
3. When it finishes, run `npx agentmash@latest status` in the repository we are in, and tell me in one or two plain sentences what it says: who I am signed in as, and whether this repository is connected.
4. If this repository is not connected yet, tell me to connect it at https://agentmash.dev/install (I sign in with GitHub there and pick the repository; only someone who can push to it can connect it). You cannot do that step for me.
5. Tell me to restart Crush, because it loads hooks and MCP servers as it starts.

Rules:

- Do not commit, push, or change any file in this repository. The login is for this machine only: it adds agentmash's hook and MCP server to ~/.config/crush/crush.json, and a sign-in token to ~/.agentmash/.
- If a command fails, show me the error and stop; do not work around it.
- To undo everything later, the command is `npx agentmash logout`.

crush has one hook, before a tool runs, and shows a hook’s note to the agent only after the edit, so agentmash refuses the agent’s first edit of a file a teammate is in, with the warning as the reason, once per session and file every ten minutes; the next try goes through. crush has no hook after an edit or for a prompt: an edit is reported once its file changes, and prompts, sessions and shell commands are not.

checked : warns by pausing the first edit and edits only, once the file changes: from our tests against recorded payloads, not run live yet; mcp tools: from crush’s docs, not tested yet.

agentmash inside factory droid

  1. connect the repository, once

    someone who can push to it signs in with github at agentmash.dev/install and picks it. nothing is committed to the repository.

  2. sign this machine in

    it finds droid and adds agentmash’s hooks to ~/.factory/hooks.json and its mcp server to ~/.factory/mcp.json. hooks you keep in ~/.factory/settings.json are copied into hooks.json beside them, since droid reads that file in their place.

    npx agentmash login

  3. restart droid

    droid reads its hooks as a session starts. if it says they changed, review them in /hooks.

or paste this into factory droid

Set up agentmash on this machine for me, for Factory Droid. agentmash lets my team's AI coding agents see what each other are working on. Once it is set up, in a repository my team has connected: my prompts and your edits show up in my team's live room, and before you edit a file a teammate's agent just changed, agentmash stops your first try and says who changed it, what changed and why; your next try goes through.

Do this:

1. Check that Node.js 18.12 or newer is installed (`node --version`). If it is not, stop and tell me how to install it.
2. Run `npx agentmash@latest login` with a timeout of at least 10 minutes. It needs the network and writes to ~/.factory and ~/.agentmash, outside this repository, so if Droid's sandbox stops it, ask me to let it run outside the sandbox. It prints a link and a code, opens my browser, and waits while I sign in with GitHub. Show me the link and the code, and tell me to finish signing in there. Do not try to sign in for me.
3. When it finishes, run `npx agentmash@latest status` in the repository we are in, and tell me in one or two plain sentences what it says: who I am signed in as, and whether this repository is connected.
4. If this repository is not connected yet, tell me to connect it at https://agentmash.dev/install (I sign in with GitHub there and pick the repository; only someone who can push to it can connect it). You cannot do that step for me.
5. Tell me to restart Droid, so it loads the hooks, and to review them in /hooks if it says they changed.

Rules:

- Do not commit, push, or change any file in this repository. The login is for this machine only: it adds agentmash's hooks to ~/.factory/hooks.json, its MCP server to ~/.factory/mcp.json, and a sign-in token to ~/.agentmash/.
- If a command fails, show me the error and stop; do not work around it.
- To undo everything later, the command is `npx agentmash logout`.

droid can’t add a note to an edit it lets through, so agentmash refuses the agent’s first edit of a file a teammate is in, with the warning as the reason, once per session and file every ten minutes; the next try goes through. built from factory’s docs and recorded payloads, not yet run against a live droid.

checked : warns by pausing the first edit and reports edits & prompts: from our tests against recorded payloads, not run live yet; mcp tools: from factory droid’s docs, not tested yet.

agentmash inside gemini cli

  1. connect the repository, once

    someone who can push to it signs in with github at agentmash.dev/install and picks it. nothing is committed to the repository.

  2. sign this machine in

    it finds the gemini command and adds agentmash’s hooks and its mcp server to ~/.gemini/settings.json, after your own. a settings file with comments in it is left as it is: take them out and run login again.

    npx agentmash login

  3. trust the folder, then restart gemini cli

    gemini cli runs hooks and mcp servers only in folders you trust. when it asks about your repository, trust it.

or paste this into gemini cli

Set up agentmash on this machine for me, for Gemini CLI. agentmash lets my team's AI coding agents see what each other are working on. Once it is set up, in a repository my team has connected: my prompts and your edits show up in my team's live room, and before you edit a file a teammate's agent just changed, agentmash stops your first try and says who changed it, what changed and why; your next try goes through.

Do this:

1. Check that Node.js 18.12 or newer is installed (`node --version`). If it is not, stop and tell me how to install it.
2. Run `npx agentmash@latest login` with a timeout of at least 10 minutes: it prints a link and a code, opens my browser, and waits while I sign in with GitHub. Show me the link and the code, and tell me to finish signing in there. Do not try to sign in for me.
3. When it finishes, run `npx agentmash@latest status` in the repository we are in, and tell me in one or two plain sentences what it says: who I am signed in as, and whether this repository is connected.
4. If this repository is not connected yet, tell me to connect it at https://agentmash.dev/install (I sign in with GitHub there and pick the repository; only someone who can push to it can connect it). You cannot do that step for me.
5. Tell me to restart Gemini CLI, and that it runs agentmash only in folders I trust: when it asks whether to trust this folder, I say yes. If the login said ~/.gemini/settings.json has comments in it, tell me that too, and do not edit that file yourself.

Rules:

- Do not commit, push, or change any file in this repository. The login is for this machine only: it adds agentmash's hooks and MCP server to ~/.gemini/settings.json, and a sign-in token to ~/.agentmash/.
- If a command fails, show me the error and stop; do not work around it.
- To undo everything later, the command is `npx agentmash logout`.

gemini cli can’t add a warning to an edit it lets through, so agentmash refuses the agent’s first edit of a file a teammate is in, with the warning as the reason, once per session and file every ten minutes; the next try goes through. it runs hooks and mcp servers only in folders it trusts. a shell command (sed, a heredoc) is not warned about before it runs; the files it changed are found in the working tree afterwards and reported.

checked : warns by pausing the first edit and reports edits & prompts: from our tests against recorded payloads, not run live yet; mcp tools: from gemini cli’s docs, not tested yet.

agentmash inside goose

  1. connect the repository, once

    someone who can push to it signs in with github at agentmash.dev/install and picks it. nothing is committed to the repository.

  2. sign this machine in

    it finds goose and adds a plugin folder of agentmash’s, ~/.agents/plugins/agentmash, with its hooks and its mcp server. goose’s own config.yaml is left as it is.

    npx agentmash login

  3. start a new goose session

    the goose cli, goose desktop and goose acp all pick it up in the next session they start; a session already open keeps what it started with.

or paste this into goose

Set up agentmash on this machine for me, for goose. agentmash lets my team's AI coding agents see what each other are working on. Once it is set up, in a repository my team has connected: my prompts and your edits show up in my team's live room, and before you edit a file a teammate's agent just changed, agentmash stops your first try and says who changed it, what changed and why; goose calls that a policy denial, but your next try goes through.

Do this:

1. Check that Node.js 18.12 or newer is installed (`node --version`). If it is not, stop and tell me how to install it.
2. Run `npx agentmash@latest login` with a timeout of at least 10 minutes: it prints a link and a code, opens my browser, and waits while I sign in with GitHub. Show me the link and the code, and tell me to finish signing in there. Do not try to sign in for me.
3. When it finishes, run `npx agentmash@latest status` in the repository we are in, and tell me in one or two plain sentences what it says: who I am signed in as, and whether this repository is connected.
4. If this repository is not connected yet, tell me to connect it at https://agentmash.dev/install (I sign in with GitHub there and pick the repository; only someone who can push to it can connect it). You cannot do that step for me.
5. Tell me to start a new goose session: goose reads agentmash's hooks when a session starts.

Rules:

- Do not commit, push, or change any file in this repository. The login is for this machine only: it adds a plugin folder, ~/.agents/plugins/agentmash, with agentmash's hooks and MCP server, and a sign-in token to ~/.agentmash/.
- If a command fails, show me the error and stop; do not work around it.
- To undo everything later, the command is `npx agentmash logout`.

goose reads nothing but a decision before a tool, so agentmash refuses the agent’s first edit of a file a teammate is in, with the warning as the reason, once per session and file every ten minutes; goose calls that a policy denial, and the next try goes through. a shell command names no file: what it changed is found in the working tree afterwards and reported. goose desktop sends no session end, and goose’s subagents run no hooks.

checked : warns by pausing the first edit and reports edits & prompts: from our tests against recorded payloads, not run live yet; mcp tools: from goose’s docs, not tested yet.

agentmash inside junie

  1. connect the repository, once

    someone who can push to it signs in with github at agentmash.dev/install and picks it. nothing is committed to the repository.

  2. sign this machine in

    it finds junie’s cli and adds agentmash’s hooks to ~/.junie/config.json and its mcp server to ~/.junie/mcp/mcp.json.

    npx agentmash login

  3. start a new junie session

    junie reads its hooks as a session starts. it asks before each agentmash mcp tool call unless you allow it; the login adds no allow rule.

or paste this into junie

Set up agentmash on this machine for me, for Junie. agentmash lets my team's AI coding agents see what each other are working on. Once it is set up, in a repository my team has connected: my prompts and your edits show up in my team's live room, and before you edit a file a teammate's agent just changed, you are told who changed it, what changed and why.

Do this:

1. Check that Node.js 18.12 or newer is installed (`node --version`). If it is not, stop and tell me how to install it.
2. Run `npx agentmash@latest login` with a timeout of at least 10 minutes. It needs the network and writes to ~/.junie and ~/.agentmash, outside this repository. It prints a link and a code, opens my browser, and waits while I sign in with GitHub. Show me the link and the code, and tell me to finish signing in there. Do not try to sign in for me.
3. When it finishes, run `npx agentmash@latest status` in the repository we are in, and tell me in one or two plain sentences what it says: who I am signed in as, and whether this repository is connected.
4. If this repository is not connected yet, tell me to connect it at https://agentmash.dev/install (I sign in with GitHub there and pick the repository; only someone who can push to it can connect it). You cannot do that step for me.
5. Tell me to start a new Junie session, because Junie reads its hooks as a session starts.

Rules:

- Do not commit, push, or change any file in this repository. The login is for this machine only: it adds agentmash's hooks to ~/.junie/config.json, its MCP server to ~/.junie/mcp/mcp.json, and a sign-in token to ~/.agentmash/.
- Do not change ~/.junie/allowlist.json or turn on brave mode for me.
- If a command fails, show me the error and stop; do not work around it.
- To undo everything later, the command is `npx agentmash logout`.

junie has no hook after a tool runs, so an edit shows up in the room when the agent’s next edit starts or its task ends, and a cancelled task’s edits at the next prompt; a shell command’s edits show up at the next prompt. prompts are reported in junie’s terminal ui, not in batch runs. junie’s docs still call hooks early access, and junie in the ide runs none.

checked : reports an edit a step later: from our tests against recorded payloads, not run live yet; warns before an edit (to be verified) and asks you before each call: from junie’s docs, not tested yet.

agentmash inside kilo code

  1. connect the repository, once

    someone who can push to it signs in with github at agentmash.dev/install and picks it. nothing is committed to the repository.

  2. sign this machine in

    it finds kilo code, writes agentmash’s plugin to ~/.config/kilo/agentmash.js, and names it, with its mcp server, in kilo’s global config, ~/.config/kilo/kilo.jsonc.

    npx agentmash login

  3. restart kilo code

    the cli, the vs code extension and the jetbrains plugin run the same backend, so this covers all three. in an editor, reload the window.

or paste this into kilo code

Set up agentmash on this machine for me, for Kilo Code. agentmash lets my team's AI coding agents see what each other are working on. Once it is set up, in a repository my team has connected: my prompts and your edits show up in my team's live room, and before you edit a file a teammate's agent just changed, agentmash stops your first try and says who changed it, what changed and why; your next try goes through.

Do this:

1. Check that Node.js 18.12 or newer is installed (`node --version`). If it is not, stop and tell me how to install it.
2. Run `npx agentmash@latest login` with a timeout of at least 10 minutes: it prints a link and a code, opens my browser, and waits while I sign in with GitHub. Show me the link and the code, and tell me to finish signing in there. Do not try to sign in for me.
3. When it finishes, run `npx agentmash@latest status` in the repository we are in, and tell me in one or two plain sentences what it says: who I am signed in as, and whether this repository is connected.
4. If this repository is not connected yet, tell me to connect it at https://agentmash.dev/install (I sign in with GitHub there and pick the repository; only someone who can push to it can connect it). You cannot do that step for me.
5. Tell me to restart Kilo Code (in VS Code or a JetBrains IDE, reload the window), because it loads plugins as it starts.

Rules:

- Do not commit, push, or change any file in this repository. The login is for this machine only: it writes agentmash's plugin to ~/.config/kilo/agentmash.js, names it and its MCP server in Kilo's global config (~/.config/kilo/kilo.jsonc), and puts a sign-in token in ~/.agentmash/.
- If a command fails, show me the error and stop; do not work around it.
- To undo everything later, the command is `npx agentmash logout`.

kilo code’s plugins can’t add a warning to an edit they let through, so agentmash refuses the agent’s first edit of a file a teammate is in, with the warning as the reason, once per session and file every ten minutes; the next try goes through. a shell command (sed, a heredoc) names no file, so nothing warns before it; the files it changed are found in the working tree afterwards and reported. with KILO_PURE=1 kilo loads no plugins, this one included.

checked : warns by pausing the first edit and reports edits & prompts: tested live; mcp tools: from kilo code’s docs, not tested yet.

agentmash inside kiro

  1. connect the repository, once

    someone who can push to it signs in with github at agentmash.dev/install and picks it. nothing is committed to the repository.

  2. sign this machine in

    it finds kiro and adds agentmash’s hooks to ~/.kiro/hooks/agentmash.json and its mcp server to ~/.kiro/settings/mcp.json.

    npx agentmash login

  3. use the ide, or kiro-cli on v3

    the kiro ide runs the hooks from its next chat. kiro-cli runs them on its v3 engine: start it with kiro-cli --v3, or pick v3 in its chat.agentEngine setting. on its default engine it doesn’t.

or paste this into kiro

Set up agentmash on this machine for me, for Kiro. agentmash lets my team's AI coding agents see what each other are working on. Once it is set up, in a repository my team has connected: my prompts and your edits show up in my team's live room, and before you edit a file a teammate's agent just changed, agentmash stops your first try and says who changed it, what changed and why; your next try goes through.

Do this:

1. Check that Node.js 18.12 or newer is installed (`node --version`). If it is not, stop and tell me how to install it.
2. Run `npx agentmash@latest login` with a timeout of at least 10 minutes: it prints a link and a code, opens my browser, and waits while I sign in with GitHub. Show me the link and the code, and tell me to finish signing in there. Do not try to sign in for me. If Kiro will not let the command write to ~/.kiro, ask me to run it in a terminal myself.
3. When it finishes, run `npx agentmash@latest status` in the repository we are in, and tell me in one or two plain sentences what it says: who I am signed in as, and whether this repository is connected.
4. If this repository is not connected yet, tell me to connect it at https://agentmash.dev/install (I sign in with GitHub there and pick the repository; only someone who can push to it can connect it). You cannot do that step for me.
5. Tell me that the Kiro IDE runs agentmash's hooks from its next chat, and that kiro-cli runs them only on its V3 engine (`kiro-cli --v3`).

Rules:

- Do not commit, push, or change any file in this repository. The login is for this machine only: it adds agentmash's hooks to ~/.kiro/hooks/agentmash.json, its MCP server to ~/.kiro/settings/mcp.json, and a sign-in token to ~/.agentmash/.
- If a command fails, show me the error and stop; do not work around it.
- To undo everything later, the command is `npx agentmash logout`.

kiro shows its agent nothing a hook says before an edit unless the hook refuses it, so agentmash refuses the agent’s first edit of a file a teammate is in, with the warning as the reason, once per conversation and file every ten minutes; the next try goes through. kiro shows a card in the chat for each hook it runs. kiro-cli on its default engine doesn’t run these hooks, so there it’s seen through the git hook only.

checked : warns by pausing the first edit and reports edits & prompts: from our tests against recorded payloads, not run live yet; mcp tools: from kiro’s docs, not tested yet.

agentmash inside opencode

  1. connect the repository, once

    someone who can push to it signs in with github at agentmash.dev/install and picks it. nothing is committed to the repository.

  2. sign this machine in

    it finds opencode and adds agentmash’s plugin to ~/.config/opencode/plugin/agentmash.js and its mcp server to opencode’s global config in ~/.config/opencode/.

    npx agentmash login

  3. restart opencode

    opencode loads plugins only as it starts.

or paste this into opencode

Set up agentmash on this machine for me, for OpenCode. agentmash lets my team's AI coding agents see what each other are working on. Once it is set up, in a repository my team has connected: my prompts and your edits show up in my team's live room, and before you edit a file a teammate's agent just changed, agentmash stops your first try and says who changed it, what changed and why; your next try goes through.

Do this:

1. Check that Node.js 18.12 or newer is installed (`node --version`). If it is not, stop and tell me how to install it.
2. Run `npx agentmash@latest login` with a timeout of at least 10 minutes (600000 ms): it prints a link and a code, opens my browser, and waits while I sign in with GitHub. Show me the link and the code, and tell me to finish signing in there. Do not try to sign in for me.
3. When it finishes, run `npx agentmash@latest status` in the repository we are in, and tell me in one or two plain sentences what it says: who I am signed in as, and whether this repository is connected.
4. If this repository is not connected yet, tell me to connect it at https://agentmash.dev/install (I sign in with GitHub there and pick the repository; only someone who can push to it can connect it). You cannot do that step for me.
5. Tell me to restart OpenCode, because it loads plugins only as it starts.

Rules:

- Do not commit, push, or change any file in this repository. The login is for this machine only: it adds agentmash's plugin to ~/.config/opencode/plugin/agentmash.js, its MCP server to OpenCode's global config in ~/.config/opencode/, and a sign-in token to ~/.agentmash/.
- If a command fails, show me the error and stop; do not work around it.
- To undo everything later, the command is `npx agentmash logout`.

opencode can’t add a warning to an edit it lets through, so agentmash refuses the agent’s first edit of a file a teammate is in, with the warning as the reason, once per session and file every ten minutes; the next try goes through. a shell command (sed, a heredoc) names no file, so nothing warns before it; what it changed is reported afterwards. opencode --pure runs no plugins, agentmash’s included.

checked : warns by pausing the first edit and reports edits & prompts: tested live; mcp tools: from opencode’s docs, not tested yet.

agentmash inside openhands

  1. connect the repository, once

    someone who can push to it signs in with github at agentmash.dev/install and picks it. nothing is committed to the repository.

  2. sign this machine in

    it finds openhands and adds agentmash’s hooks to ~/.openhands/hooks.json and its mcp server to ~/.openhands/mcp.json for the openhands cli, and a plugin of its own, ~/.openhands/plugins/agentmash-openhands, for agent canvas.

    npx agentmash login

  3. start a new conversation

    openhands reads hooks and plugins as a conversation starts. in the cli, a repository with its own .openhands/hooks.json runs only those.

or paste this into openhands

Set up agentmash on this machine for me, for OpenHands. agentmash lets my team's AI coding agents see what each other are working on. Once it is set up, in a repository my team has connected: my prompts and your edits show up in my team's live room, and before you edit a file a teammate's agent just changed, agentmash stops your first try and says who changed it, what changed and why; your next try goes through.

Do this:

1. Check that Node.js 18.12 or newer is installed (`node --version`). If it is not, stop and tell me how to install it.
2. Run `npx agentmash@latest login` with a timeout of at least 10 minutes: it prints a link and a code, opens my browser, and waits while I sign in with GitHub. Show me the link and the code, and tell me to finish signing in there. Do not try to sign in for me.
3. When it finishes, run `npx agentmash@latest status` in the repository we are in, and tell me in one or two plain sentences what it says: who I am signed in as, and whether this repository is connected.
4. If this repository is not connected yet, tell me to connect it at https://agentmash.dev/install (I sign in with GitHub there and pick the repository; only someone who can push to it can connect it). You cannot do that step for me.
5. Tell me to start a new OpenHands conversation, because OpenHands reads hooks and plugins as a conversation starts.

Rules:

- Do not commit, push, or change any file in this repository. The login is for this machine only: it adds agentmash's hooks to ~/.openhands/hooks.json, its MCP server to ~/.openhands/mcp.json, a plugin folder under ~/.openhands/plugins for Agent Canvas, and a sign-in token to ~/.agentmash/.
- If a command fails, show me the error and stop; do not work around it.
- To undo everything later, the command is `npx agentmash logout`.

openhands reads nothing a hook adds to a call it lets run, so agentmash refuses the agent’s first edit of a file a teammate is in, with the warning as the reason, once per conversation and file every ten minutes; the next try goes through. in the cli’s confirm-every-action mode you’re asked about that edit first: confirm it, and the agent reads the warning. an edit made in the terminal names no file, so nothing warns before it; the working tree reports it afterwards. the openhands cli is no longer maintained; agent canvas runs agentmash as a plugin on its local agent server, not in docker or on a remote one.

checked : warns by pausing the first edit and reports edits & prompts: from our tests against recorded payloads, not run live yet; mcp tools: from openhands’s docs, not tested yet.

agentmash inside qwen code

  1. connect the repository, once

    someone who can push to it signs in with github at agentmash.dev/install and picks it. nothing is committed to the repository.

  2. sign this machine in

    it finds qwen code and adds agentmash’s hooks and its mcp server to ~/.qwen/settings.json (or $QWEN_HOME/settings.json), after yours.

    npx agentmash login

  3. restart qwen code

    qwen code reads its hooks as it starts. there is nothing to approve: hooks in your own settings run in every folder.

or paste this into qwen code

Set up agentmash on this machine for me, for Qwen Code. agentmash lets my team's AI coding agents see what each other are working on. Once it is set up, in a repository my team has connected: my prompts and your edits show up in my team's live room, and the first time you edit a file a teammate's agent just changed, that edit is stopped once with who changed it, what changed and why; the same edit again goes through.

Do this:

1. Check that Node.js 18.12 or newer is installed (`node --version`). If it is not, stop and tell me how to install it.
2. Run `npx agentmash@latest login` with a timeout of at least 10 minutes. It needs the network and writes to ~/.qwen and ~/.agentmash, outside this repository, so if the sandbox stops it, ask me to let it run outside the sandbox. It prints a link and a code, opens my browser, and waits while I sign in with GitHub. Show me the link and the code, and tell me to finish signing in there. Do not try to sign in for me.
3. When it finishes, run `npx agentmash@latest status` in the repository we are in, and tell me in one or two plain sentences what it says: who I am signed in as, and whether this repository is connected.
4. If this repository is not connected yet, tell me to connect it at https://agentmash.dev/install (I sign in with GitHub there and pick the repository; only someone who can push to it can connect it). You cannot do that step for me.
5. Tell me to restart Qwen Code, because it reads its hooks as it starts.

Rules:

- Do not commit, push, or change any file in this repository. The login is for this machine only: it adds agentmash's hooks and MCP server to ~/.qwen/settings.json, and a sign-in token to ~/.agentmash/.
- If a command fails, show me the error and stop; do not work around it.
- To undo everything later, the command is `npx agentmash logout`.

qwen code runs a hook before an edit only after the edit is approved, and can’t pass it a warning without stopping it: the first edit of a file a teammate’s agent is in is refused once, with the warning as the reason, and the same edit again goes through (in its default mode you approve that edit twice). a shell command names no file; the files it changed are found in the working tree afterwards. prompts are reported from 0.21.1, and only where qwen code says which prompt you typed (an ide that runs it over acp may not).

checked : refuses the first edit once, with the warning as the reason and reports edits & prompts: from our tests against recorded payloads, not run live yet; mcp tools: from qwen code’s docs, not tested yet.

agentmash inside continue

  1. connect the repository, once

    someone who can push to it signs in with github at agentmash.dev/install and picks it. nothing is committed to the repository.

  2. sign this machine in

    it adds agentmash’s mcp server to mcpServers in ~/.continue/config.yaml, which the vs code and jetbrains extensions and cn all read, after your own servers. it only adds to a config.yaml you already have: open continue once first.

    npx agentmash login

  3. reload continue

    start a new cn, or reload your editor window. continue asks before each mcp tool call.

or paste this into continue

Set up agentmash on this machine for me, for Continue. agentmash lets my team's AI coding agents see what each other are working on. Continue runs no hooks, so agentmash gives it MCP tools only: once it is set up, in a repository my team has connected, you can ask agentmash who is working on what, and whether anyone touched a file, before you edit it. Your edits do not show in my team's room.

Do this:

1. Check that Node.js 18.12 or newer is installed (`node --version`). If it is not, stop and tell me how to install it.
2. Run `npx agentmash@latest login` with a timeout of at least 10 minutes: it prints a link and a code, opens my browser, and waits while I sign in with GitHub. Show me the link and the code, and tell me to finish signing in there. Do not try to sign in for me.
3. When it finishes, run `npx agentmash@latest status` in the repository we are in, and tell me in one or two plain sentences what it says: who I am signed in as, and whether this repository is connected.
4. If this repository is not connected yet, tell me to connect it at https://agentmash.dev/install (I sign in with GitHub there and pick the repository; only someone who can push to it can connect it). You cannot do that step for me.
5. If the login said it skipped Continue, tell me why in one sentence. Otherwise tell me to start a new cn or reload my editor, and that Continue asks before each agentmash tool call.

Rules:

- Do not commit, push, or change any file in this repository. The login is for this machine only: it adds agentmash's MCP server to ~/.continue/config.yaml, and a sign-in token to ~/.agentmash/.
- If a command fails, show me the error and stop; do not work around it.
- To undo everything later, the command is `npx agentmash logout`.

continue’s makers have stopped work on it (2.0.0 was the last release). it runs no hooks, so there’s no warning before an edit and its edits don’t show in the room: it gets the mcp tools. a config.yaml login can’t edit safely, or the older config.json, is left alone and login says why. the git hook reports its commits.

checked : from continue’s docs, not tested yet.

agentmash inside zed

  1. connect the repository, once

    someone who can push to it signs in with github at agentmash.dev/install and picks it. nothing is committed to the repository.

  2. sign this machine in

    it finds zed and adds agentmash’s mcp server to context_servers in ~/.config/zed/settings.json, after your own settings, comments kept. zed picks it up without a restart.

    npx agentmash login

  3. allow the tools in zed’s agent

    zed asks before each mcp tool call: choose “always” for agentmash’s tools if you want the agent to check without asking. zed’s own agent has no hooks, so it hears about teammates only when it asks.

or paste this into zed

Set up agentmash on this machine for me, for Zed. agentmash lets my team's AI coding agents see what each other are working on. Zed's own agent has no hooks, so agentmash gives it MCP tools only: once it is set up, in a repository my team has connected, you can ask agentmash who is working on what, and whether anyone touched a file, before you edit it. Your edits do not show in my team's room.

Do this:

1. Check that Node.js 18.12 or newer is installed (`node --version`). If it is not, stop and tell me how to install it.
2. Run `npx agentmash@latest login` with a timeout of at least 10 minutes: it prints a link and a code, opens my browser, and waits while I sign in with GitHub. Show me the link and the code, and tell me to finish signing in there. Do not try to sign in for me.
3. When it finishes, run `npx agentmash@latest status` in the repository we are in, and tell me in one or two plain sentences what it says: who I am signed in as, and whether this repository is connected.
4. If this repository is not connected yet, tell me to connect it at https://agentmash.dev/install (I sign in with GitHub there and pick the repository; only someone who can push to it can connect it). You cannot do that step for me.
5. Tell me that Zed asks before each agentmash tool call, and that I can choose Always for those tools on the approval card.

Rules:

- Do not commit, push, or change any file in this repository. The login is for this machine only: it adds agentmash's MCP server to context_servers in my Zed settings (~/.config/zed/settings.json), and a sign-in token to ~/.agentmash/.
- If a command fails, show me the error and stop; do not work around it.
- To undo everything later, the command is `npx agentmash logout`.

zed’s agent has no hooks (checked in its source at 1.22), so there’s no warning before an edit and its edits don’t show in the room: it gets the mcp tools and asks when it chooses to. zed’s claude agent threads run claude code’s own hooks. the git hook reports zed’s commits.

checked : from zed’s docs, not tested yet.

agentmash inside any editor

  1. set the repository up with init, once

    the git hook reads its room from what npx agentmash init commits, not from a login: someone who can push to the repository runs this at its root and commits what it writes. a repository connected on the website already keeps its room: init there stops and says how to commit one anyway.

    npx agentmash init

  2. add the git hook to each checkout

    it reports each commit’s files, subject and line counts, in a repository set up with npx agentmash init: the hook reads its room from the config init commits, not from a login.

    npx agentmash git-hook install

for vim, emacs, aider, zed or anything else. it shows where someone landed, not what they’re doing: no warning before an edit, and nothing until a commit.

checked : tested live.

agentmash inside your own tool

  1. connect the repository, once

    someone who can push to it signs in with github at agentmash.dev/install and picks it. nothing is committed to the repository.

  2. sign this machine in

    it puts agentmash’s generic hook in ~/.agentmash/hooks, beside the scripts every agent in the other tabs runs, and the mcp server with it.

    npx agentmash login

  3. print the line for each hook

    one line for each event your tool has a hook for (pre_edit, post_edit, prompt, session_start, stop, session_end), answering in the format it reads: --dialect claude, deny-once, plain or json. register the node line it prints, never npx. the public api has the input it reads.

    npx agentmash hook --print-command --client mytool --event pre_edit

  4. or call the api

    a tool that speaks http makes a personal token and calls /api/v1: GET /api/v1/check gives the advice to show before an edit, POST /api/v1/events reports one. the public api has every route.

    npx agentmash token create --name mytool

agentmash connects the agents in the other tabs by itself. for a tool it has no connector for, the generic hook is the same script they run and the api reaches the same room: what your agent is told before an edit depends on what its hooks can say, and the dialect you pick says it.

checked : from our tests against recorded payloads, not run live yet.

what each step does

npx agentmash login
on each developer’s machine, once. it signs the machine in with github and connects the coding agents it finds, in their own settings on that machine. nothing goes into a repository.
agentmash.dev/install
each repository, once, by someone who can push to it. it is in the browser, with a github sign-in of its own: signing in from the terminal doesn’t sign the browser in. nothing is committed.
npx agentmash init optional
the only step that commits. it writes agentmash’s hooks and its mcp server’s entry into the repository, so teammates get claims and the mcp tools in claude code with a git pull, and the git hook has a room to report to. on a signed-in machine it connects the repository too. a repository connected already keeps its room: there, init stops and says how to commit one anyway.

agentmash and other agents

  • jetbrains aimcp only, set up in its settings
  • aiderthe git hook

any of them can report at commit time through the git hook, and a tool with hooks of its own can run the generic hook. missing yours? hello@agentmash.dev

to take it all back out: npx agentmash logout.